A Deep Ensemble Learning Approach Based on a Vision Transformer and Neural Network for Multi-Label Image Classification

Big Data and Cognitive Computing, Journal Year: 2025, Volume and Issue: 9(2), P. 39 - 39

Published: Feb. 11, 2025

Convolutional Neural Networks (CNNs) have proven to be very effective in image classification due their status as a powerful feature learning algorithm. Traditional approaches considered the problem of multiclass classification, where goal is classify set objects at once. However, co-occurrence can make discriminative features target less salient and may lead overfitting model, resulting lower performance. To address this, we propose multi-label ensemble model including Vision Transformer (ViT) CNN for directly detecting one or multiple an image. First, improve MobileNetV2 DenseNet201 models using extra convolutional layers strengthen classification. In detail, three convolution are applied parallel end both models. ViT learn dependencies among distant positions local making it tool Finally, algorithm used combine predictions ViT, modified MobileNetV2, bands increased accuracy voting system. The performance proposed examined on four benchmark datasets, achieving accuracies 98.24%, 98.89%, 99.91%, 96.69% ASCAL VOC 2007, PASCAL 2012, MS-COCO, NUS-WIDE 318, respectively, showing that our framework enhance current state-of-the-art methods.

Language: Английский

A Transformer-Based Symmetric Diffusion Segmentation Network for Wheat Growth Monitoring and Yield Counting

Agriculture, Journal Year: 2025, Volume and Issue: 15(7), P. 670 - 670

Published: March 21, 2025

A wheat growth and counting analysis model based on instance segmentation is proposed in this study to address the challenges of monitoring yield prediction high-density agricultural environments. The integrates transformer architecture with a symmetric attention mechanism employs diffusion module for precise measurement instances. By introducing an aggregated loss function, effectively optimizes both accuracy performance. Experimental results show that excels across several evaluation metrics. Specifically, task, using achieved Precision 0.91, Recall 0.87, Accuracy 0.89, mAP@75 0.88, F1-score significantly outperforming other baseline methods. For model’s reached 0.95, was 0.90, 0.93, 0.92, demonstrating marked advantage monitoring. Finally, provides novel effective method environments, offering substantial support future intelligent decision-making systems.

Language: Английский